Pretrained Natural Language Processing Model for Intent Recognition (BERT-IR)

Abstract

Intent Recognition (IR) is considered a key area in Natural Language Processing (NLP). It has crucial usage in various applications. One is the Search Engine- Interpreting the context of text searched by the user improves the response time and helps the search engines give appropriate outputs. Another can be Social Media Analytics-Analysing profiles of users on different social media platforms has become a necessity in today’s applications like recommendation systems in the online world, digital marketing, and a lot more. Many researchers are using different techniques for achieving intent recognition but getting high accuracy in intent recognition is crucial. In this work, named BERT-IR, a pre-trained Natural Language Processing model called as BERT model, along with few add-ons, is applied for the task of Intent Recognition. We have achieved an accuracy of 97.67% on a widely used dataset which shows the capability and efficiency of our work. For comparison purposes, we have applied primarily used Machine Learning techniques, namely Naive Bayes, Logistic Regression, Decision Tree, Random Forest, and Gradient Boost as well as Deep Learning Techniques used for intent recognition like Recurrent Neural Network, Long Short Term Memory Network, and Bidirectional Long Short Term Memory Network on the same dataset and evaluated the accuracy. It is found out that BERT-IR’s accuracy is far better than that of the other models implemented.